Efficient Tennis Player Tracking Using Online Video Highlights

In tennis game analysis, player tracking serves as a foundational task before any advanced analysis. However, current methods encounter numerous challenges. The most accessible resources are tennis videos from the internet, which often include post-processed shots, posing challenges for frame-based tracking. Moreover, there is significant potential for enhancing tracking performance by striking a balance between accuracy and efficiency. In this paper, we present a holistic solution for efficient player tracking utilizing online tennis videos. For each video, we employ a lightweight image classifier as a pre-filter to bypass undesired post-processed frames. In player tracking, leveraging the distinctive movement patterns of tennis players, we introduce a novel tracking method, called T-tracker, aims to streamline calculations by eliminating unnecessary computations while maintaining the accuracy. We build up the corresponding baselines to evaluate the efficiency. The results show that the frequency of ID switch is reduced by 68.85% in “Rally” frames, and 73.11% in total when compared to SORT. Moreover, the speed for tracking is increased by 15.27% compared to DeepSORT. Additionally, the total processing speed for a 30second video sees an improvement of 2.01% compared to DeepSORT.